def kps_to_BB(kps):
"""
Determine imgaug bounding box from imgaug keypoints
"""
extend = 3 # To make the bounding box a little bit bigger
kpsx = [kp.x for kp in kps.keypoints]
minx = max(0, int(min(kpsx) - extend))
maxx = min(imgW, int(max(kpsx) + extend))
kpsy = [kp.y for kp in kps.keypoints]
miny = max(0, int(min(kpsy) - extend))
maxy = min(imgH, int(max(kpsy) + extend))
if minx == maxx or miny == maxy:
return None
else:
return ia.BoundingBox(x1=minx, y1=miny, x2=maxx, y2=maxy)
def kps_to_BB(kps):
"""
通过keypoints确定图像中的Bbox
"""
extend=3 #
kpsx=[kp.x for kp in kps.keypoints] #取每个keypoints的kp的x值
minx=max(0,int(min(kpsx)-extend))# 左x
maxx=min(imgW,int(max(kpsx)+extend))#右x
kpsy=[kp.y for kp in kps.keypoints]#取每个keypoints的kp的y值
miny=max(0,int(min(kpsy)-extend))# 左y
maxy=min(imgH,int(max(kpsy)+extend)) # 右y
if minx==maxx or miny==maxy:# 面积为0的情况--返回
return None
else:
return ia.BoundingBox(x1=minx,y1=miny,x2=maxx,y2=maxy)# 否则返回一个Bbox
def img_augmentation(augmentation, img, bbox):
"""
Augment image and bounding boxes !!
:param augmentation: augmentation settings
:param img: Only one image is needed. Not batch images
:param bbox: [[x1,y1,x2,y2],[x1,y1,x2,y2]...]
:return: Returns augment image and bbox
"""
# img_copy = img.copy()
image_shape = img.shape
h, w = image_shape[0:2]
# Convert the stochastic sequence of augmenters to a deterministic one.
# The deterministic sequence will always apply the exactly same effects to the images.
det = augmentation.to_deterministic()
img_aug = det.augment_image(img)
ia_bbox = list()
for bounding_box in bbox:
x1, y1, x2, y2 = bounding_box
ia_bbox.append(ia.BoundingBox(x1=x1, y1=y1, x2=x2, y2=y2))
bbs = ia.BoundingBoxesOnImage(ia_bbox, shape=image_shape)
bbs_aug = det.augment_bounding_boxes([bbs])[0]
# img = bbs_aug.draw_on_image(img)
after_bbox = list()
for bounding_box in bbs_aug.bounding_boxes:
bbox_list = [
bounding_box.x1_int, bounding_box.y1_int, bounding_box.x2_int,
bounding_box.y2_int
]
if bbox_list[0] >= w: bbox_list[0] = w - 1
if bbox_list[1] >= h: bbox_list[1] = h - 1
if bbox_list[2] >= w: bbox_list[2] = w - 1
if bbox_list[3] >= h: bbox_list[3] = h - 1
if bbox_list[0] == bbox_list[2] or bbox_list[1] == bbox_list[3]:
return img_augmentation(augmentation, img, bbox)
bbox_list = list(map(lambda x: max(x, 0), bbox_list))
after_bbox.append(bbox_list)
assert img_aug.shape == image_shape, "Augmentation shouldn't change image size"
return img_aug, after_bbox
def get_images_bboxes(df=None, path_to_images=PATH_TO_AUG_IMAGES, head=None):
"""
the function get a list of images and bboxes
and print them side by side
"""
# make a list of subgroups, each subgroup is of one file
data = namedtuple('data', ['file', 'object'])
grouped = df.groupby('filename')
gd_data = []
file_naems = []
for file, x in zip(grouped.groups.keys(), grouped.groups):
gd_data.append(data(file, grouped.get_group(x)))
file_naems.append(file)
filenames = []
images = []
bboxes = []
classes = []
for g in gd_data:
grouped_image = g.object
H = int(grouped_image['height'].iloc[0])
W = int(grouped_image['width'].iloc[0])
filename = grouped_image['filename'].iloc[0]
# create list of bboxes object to each image
bboxes_on_an_image = []
classes_on_an_image = []
for idx, row in grouped_image.iterrows():
x1 = int(row['xmin'])
y1 = int(row['ymin'])
x2 = int(row['xmax'])
y2 = int(row['ymax'])
bboxes_on_an_image.append(ia.BoundingBox(x1, y1, x2, y2))
classes_on_an_image.append(row['class'])
bboxes.append(ia.BoundingBoxesOnImage(bboxes_on_an_image,
shape=(H, W)))
classes.append(classes_on_an_image)
filenames.append(filename)
# open an image
path = os.path.join(path_to_images, filename)
image = cv.imread(path)
image = cv.cvtColor(image, cv.COLOR_BGR2RGB)
images.append(image)
return filenames, images, bboxes, classes
def create_sample():
while True:
num_objects = np.random.randint(MIN_OBJECTS_PER_IMAGE,
MAX_OBJECTS_PER_IMAGE + 1)
ok = False
for _ in range(MAX_TRIES):
bnd_boxes = []
for _ in range(num_objects):
bird_size = np.random.randint(MIN_SIZE, MAX_SIZE)
bird_ar = np.random.uniform(MIN_ASPECT_RATIO, MAX_ASPECT_RATIO)
bird_width = min(IMAGE_SIZE, int(bird_size * np.sqrt(bird_ar)))
bird_height = min(IMAGE_SIZE,
int(bird_size / np.sqrt(bird_ar)))
y1 = np.random.randint(0, IMAGE_SIZE - bird_height + 1)
x1 = np.random.randint(0, IMAGE_SIZE - bird_width + 1)
y2 = y1 + bird_height
x2 = x1 + bird_width
bnd_boxes.append(
ia.BoundingBox(x1=x1, y1=y1, x2=x2, y2=y2, label='bird'))
ok = True
for i, box1 in enumerate(bnd_boxes):
for j, box2 in enumerate(bnd_boxes):
if i != j and box1.intersection(box2):
ok = False
if ok: break
if ok: break
image = sample_background()
for box in bnd_boxes:
y1 = box.y1_int
x1 = box.x1_int
y2 = box.y2_int
x2 = box.x2_int
backg_patch = image[y1:y2, x1:x2]
image[y1:y2, x1:x2], nx1, ny1, nx2, ny2 = draw_object(backg_patch)
box.x1, box.x2 = nx1 + x1, nx2 + x1
box.y1, box.y2 = ny1 + y1, ny2 + y1
bnd_boxes = ia.BoundingBoxesOnImage(bnd_boxes, shape=image.shape)
return image, bnd_boxes
def may_augment_bbox(self, aug, ori_shape, bboxes):
imgaug_bboxes = []
for bbox in bboxes:
x1, y1, x2, y2 = bbox
imgaug_bboxes.append(imgaug.BoundingBox(x1=x1, y1=y1, x2=x2,
y2=y2))
imgaug_bboxes = aug.augment_bounding_boxes([
imgaug.BoundingBoxesOnImage(imgaug_bboxes, shape=ori_shape)
])[0].clip_out_of_image()
new_bboxes = []
for box in imgaug_bboxes.bounding_boxes:
new_bboxes.append(
np.array([box.x1, box.y1, box.x2, box.y2], dtype=np.float32))
return new_bboxes
def _process(self, image, label):
label = np.reshape(label, (-1, 5))
label = [list(label[row, :]) for row in range(label.shape[0])]
bbs = ia.BoundingBoxesOnImage([
ia.BoundingBox(x1=x1, y1=y1, x2=x2, y2=y2, label=class_id)
for x1, y1, x2, y2, class_id in label
],
shape=image.shape)
# 1. 数据增强
if self.enhance:
image, bbs = self._aug_images(image, bbs)
# 2. 图像resize
image, bbs = self._resize(image, bbs)
# 3. 制作yolo标签
label = self._to_yolo(bbs)
return image, label
def test_get_column_names__all_columns(self):
batch = ia.UnnormalizedBatch(
images=np.zeros((1, 2, 2, 3), dtype=np.uint8),
heatmaps=[np.zeros((2, 2, 1), dtype=np.float32)],
segmentation_maps=[np.zeros((2, 2, 1), dtype=np.int32)],
keypoints=[[(0, 0)]],
bounding_boxes=[[ia.BoundingBox(0, 0, 1, 1)]],
polygons=[[ia.Polygon([(0, 0), (1, 0), (1, 1)])]],
line_strings=[[ia.LineString([(0, 0), (1, 0)])]])
names = batch.get_column_names()
assert names == [
"images", "heatmaps", "segmentation_maps", "keypoints",
"bounding_boxes", "polygons", "line_strings"
]
def parseBoundingBoxes(annotations, imageShape): # input: annotation corresponding to a single bounding box # output: BoundingBox object boundingBoxes = [] for annotation in annotations: xTopLeft = annotation["x"] yTopLeft = annotation["y"] xBottomRight = xTopLeft + annotation["width"] yBottomLeft = yTopLeft + annotation["height"] boundingBox = ia.BoundingBox(x1=xTopLeft,y1=yTopLeft,x2=xBottomRight, y2=yBottomLeft) boundingBoxes.append(boundingBox) bbs = ia.BoundingBoxesOnImage(boundingBoxes, shape=imageShape) return(bbs)
def generate_augmentations(paths, output_dir, repeats, maximun=-1):
images = read_yolo_paths(paths)
seq = aumentation_properties()
print("Aumentanto {} x {}".format(len(images), repeats))
results = []
check_max = 0
for image in images:
check_max = check_max + 1
if check_max >= maximun and maximun != -1: break
filename = image["image"].split("/")[-1]
data = cv.imread(image["image"])
boxes = []
names = []
for line in image["boxes"]:
if len(line) != 5 : continue
name, x1, y1, x2, y2 = line
x1, x2, y1, y2 = convert_axis(data.shape,
(float(x1), float(y1), float(x2), float(y2)))
boxes.append(ia.BoundingBox(x1=x1, y1=y1, x2=x2, y2=y2))
names.append(name)
boxes = ia.BoundingBoxesOnImage(boxes, shape=data.shape)
seq_det = seq.to_deterministic()
images_aug = seq_det.augment_images([data] * repeats)
boxes_aug = seq_det.augment_bounding_boxes([boxes] * repeats)
for i in range(0, len(images_aug)):
image_aug = images_aug[i]
box_aug = boxes_aug[i]
result_filename = "aug-" + str(i) + filename
cv.imwrite(os.path.join(output_dir, result_filename), image_aug)
with open(os.path.join(output_dir, result_filename.split(".")[0] + ".txt"), "w") as yolo_txt:
j = 0
box_aug = box_aug.cut_out_of_image()
box = box_aug.bounding_boxes
for name in names:
#print(box[j].is_fully_within_image(data.shape))
if box[j].is_fully_within_image(data.shape):
x1, y1, x2, y2 = convert_yolo(data.shape,
(box[j].x1, box[j].x2, box[j].y1 ,box[j].y2))
yolo_txt.write("%s %.4f %.4f %.4f %.4f\n" %
(name, x1, y1, x2, y2) )
j = j + 1
i = i + 1
results.append(os.path.join(output_dir, result_filename))
shuffle_array(results)
print("Datos generados ", len(results))
return results
def cpu_augment(self, imgs, boxes):
# for bx in boxes:
# self.assert_bboxes(bx)
ia_bb = []
for n in range(len(imgs)):
c_boxes = []
for i in boxes[n]:
try:
c_boxes.append(
ia.BoundingBox(x1=i[0], y1=i[1], x2=i[2], y2=i[3]))
except AssertionError:
print('Assertion Error: ', i)
ia_bb.append(ia.BoundingBoxesOnImage(c_boxes, shape=imgs[n].shape))
seq = iaa.Sequential([
iaa.Sometimes(0.5, iaa.AddElementwise((-20, 20), per_channel=1)),
iaa.Sometimes(0.5,
iaa.AdditiveGaussianNoise(scale=(0, 0.10 * 255))),
iaa.Sometimes(0.5, iaa.Multiply((0.75, 1.25), per_channel=1)),
iaa.Sometimes(0.5, iaa.MultiplyElementwise((0.75, 1.25))),
iaa.Sometimes(0.5, iaa.GaussianBlur(sigma=(0.0, 1.0))),
iaa.Fliplr(0.5),
iaa.Sometimes(
0.95,
iaa.SomeOf(1, [
iaa.CoarseDropout(p=(0.10, 0.25),
size_percent=(0.25, 0.5)),
iaa.CoarseDropout(p=(0.0, 0.15), size_percent=(0.1, 0.25)),
iaa.Dropout(p=(0, 0.25)),
iaa.CoarseSaltAndPepper(p=(0, 0.25),
size_percent=(0.1, 0.2))
])),
iaa.Affine(scale=iap.Choice(
[iap.Uniform(0.4, 1), iap.Uniform(1, 3)]),
rotate=(-180, 180))
])
seq_det = seq.to_deterministic()
image_b_aug = seq_det.augment_images(imgs)
bbs_b_aug = seq_det.augment_bounding_boxes(ia_bb)
bbs_b_aug = [
b.remove_out_of_image().cut_out_of_image() for b in bbs_b_aug
]
return image_b_aug, [
np.array([self.bbox_r(j) for j in i.bounding_boxes])
for i in bbs_b_aug
]
def gen_batches(files, bs=5, scale=4.5, must_rotate=True):
from imgaug.augmentables.batches import UnnormalizedBatch
batches = []
trees = []
for xml_file in files:
tree = ET.parse(xml_file)
root = tree.getroot()
img = root.find('path').text
os.path.join(*img.split("\\"))
raw_img = Image.open(img)
clean(raw_img)
raw_img = ImageOps.exif_transpose(raw_img)
# Reduce image size
if scale > 1:
resize_image(raw_img, root, scale)
img_array = np.array(raw_img)
bbs = [
ia.BoundingBox(int(member[4][0].text), int(member[4][1].text),
int(member[4][2].text), int(member[4][3].text))
for member in root.findall('object')
]
# Rotated
if must_rotate:
img_aug, bbs_aug = aug_by_value_list([img_array], [bbs],
fit_output=True,
rotate=MUST_ROTATE)
else:
img_aug, bbs_aug = [img_array], [bbs]
# img_aug, bbs_aug = aug_by_value_list(img_aug, bbs_aug, scale=MUST_SCALE)
# img_aug, bbs_aug = [], []
# Original
# img_aug.insert(0, img_array)
# bbs_aug.insert(0, bbs)
images = [
img_aug_array for img_aug_array in img_aug for _ in range(bs)
]
batches.append(
UnnormalizedBatch(images=images,
bounding_boxes=[
bbs_aug_array for bbs_aug_array in bbs_aug
for _ in range(bs)
]))
trees.append(tree)
return batches, trees
def _measure_uniformity(cls, image, patch_size=5, n_patches=100):
pshalf = (patch_size - 1) // 2
image_f32 = image.astype(np.float32)
grad_x = image_f32[:, 1:] - image_f32[:, :-1]
grad_y = image_f32[1:, :] - image_f32[:-1, :]
grad = np.abs(grad_x[1:, :] + grad_y[:, 1:])
points_y = np.random.randint(0, image.shape[0], size=(n_patches, ))
points_x = np.random.randint(0, image.shape[0], size=(n_patches, ))
stds = []
for y, x in zip(points_y, points_x):
bb = ia.BoundingBox(x1=x - pshalf,
y1=y - pshalf,
x2=x + pshalf,
y2=y + pshalf)
patch = bb.extract_from_image(grad)
stds.append(np.std(patch))
return 1 / (1 + np.std(stds))
def augment_keypoints(kpts, img, d):
ia.seed(1)
#seq = iaa.Sequential([iaa.Affine(rotate=int(d)), iaa.CropAndPad(px=(-75, 0), keep_size=False)]) # rotate by exactly d degrees
seq = iaa.Sequential([
iaa.Affine(rotate=int(d)),
iaa.Affine(translate_px={
"x": -100,
"y": -100
})
])
#seq = iaa.Sequential([iaa.Affine(rotate=int(d))]) # rotate by exactly d degrees
seq_det = seq.to_deterministic()
keypoints_og = ia.KeypointsOnImage([
ia.Keypoint(x=kpts[0], y=kpts[1]),
ia.Keypoint(x=kpts[2], y=kpts[3]),
ia.Keypoint(x=kpts[4], y=kpts[5]),
ia.Keypoint(x=kpts[6], y=kpts[7])
],
shape=img.shape)
#bbs_original = ia.BoundingBoxesOnImage([ia.BoundingBox(x1=kpts[2], x2=kpts[4], y1=kpts[3], y2=kpts[5])], shape=img.shape)
#keypoints_list.append(kpts)
keypoints_aug = seq_det.augment_keypoints([keypoints_og])[0]
for i in range(len(keypoints_og.keypoints)):
before = keypoints_og.keypoints[i]
after = keypoints_aug.keypoints[i]
#print("Keypoint %d: (%.8f, %.8f) -> (%.8f, %.8f)" % (i, before.x, before.y, after.x, after.y))
print("keypoints_aug: ", keypoints_aug)
#bbs_aug = ia.BoundingBoxesOnImage([ia.BoundingBox(x1=keypoints_aug[0], x2=keypoints_aug[2], y1=keypoints_aug[1], y2=keypoints_aug[3])], shape=img.shape)
new_coords = []
for kp_idx, keypoint in enumerate(keypoints_aug.keypoints):
keypoint_old = keypoints_og.keypoints[kp_idx]
x_old, y_old = keypoint_old.x, keypoint_old.y
x_new, y_new = keypoint.x, keypoint.y
#print("[Keypoints for image #%s] before aug: x=%s y=%s | after aug: x=%s y=%s" % (img, x_old, y_old, x_new, y_new))
new_pair = (int(x_new), int(y_new))
new_coords.append(new_pair)
print("new_coords: ", new_coords)
bbs_aug = ia.BoundingBoxesOnImage([
ia.BoundingBox(x1=new_coords[0],
x2=new_coords[2],
y1=new_coords[1],
y2=new_coords[3])
],
shape=img.shape)
return new_coords, bbs_aug
def augmentImgsTest(self, plates):
augPlates = []
for plate in plates:
plateIdx = plate["plateIdx"]
plateImg = np.asarray(plate['plateImg'])
plateBoxes = plate['plateBoxes']
bbs = []
seq = iaa.Sequential([
iaa.Sometimes(0.9, iaa.GaussianBlur(sigma=(0, 0.9))),
iaa.ContrastNormalization((0.75, 2.0)),
iaa.AdditiveGaussianNoise(
loc=0, scale=(0.0, 0.2 * 255), per_channel=0.6),
iaa.Multiply((0.8, 1.2), per_channel=0.2),
iaa.Sometimes(0.8, iaa.Affine(rotate=(-2, 12),
shear=(-2, 15))),
iaa.Sometimes(0.8, iaa.Affine(shear=(-3, 9)))
],
random_order=True)
seq_det = seq.to_deterministic()
for box in plateBoxes:
bbs.append(ia.BoundingBox(box[0], box[1], box[2], box[3]))
bbsOnImage = ia.BoundingBoxesOnImage(bbs, shape=plateImg.shape)
imageAug = seq_det.augment_images([plateImg])[0]
bboxAug = seq_det.augment_bounding_boxes([bbsOnImage])[0]
bboxAug = bboxAug.remove_out_of_image().cut_out_of_image()
finalImg = Image.fromarray(imageAug)
bboxAugFormatted = []
for idx, box in enumerate(bboxAug.bounding_boxes):
bboxAugFormatted.append(
(box.x1, box.y1, box.x2, box.y2, plateBoxes[idx][4]))
augPlates.append({
"plateIdx": plateIdx,
"plateImg": Image.fromarray(imageAug),
"plateBoxes": bboxAugFormatted
})
# Visualize plate
if self.visualizePlates:
self.visualizePlate(finalImg)
return augPlates
def random_transform_group_entry(self, image: np.ndarray, annotations: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
if self._augmenter is None:
return image, annotations
seq_det = self._augmenter.to_deterministic()
image_aug = seq_det.augment_images([image])[0].astype('float32')
if annotations.shape[0] > 0:
bbs = ia.BoundingBoxesOnImage([
ia.BoundingBox(x1=a[0], y1=a[1], x2=a[2], y2=a[3]) for a in annotations
], shape=image.shape)
bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
annotations_aug = np.array([np.array([bb.x1, bb.y1, bb.x2, bb.y2, an[4]])
for (bb, an) in zip(bbs_aug.bounding_boxes, annotations)])
return image_aug, annotations_aug
else:
return image_aug, annotations
def modify_image_train(img, bb):
img_height, img_width = img.shape[0], img.shape[1]
x_min, y_min, width, height = bb
x1, y1, x2, y2 = to_imgaugbb(x_min, y_min, width, height)
bb_array = ia.BoundingBoxesOnImage([ia.BoundingBox(x1, y1, x2, y2)],
shape=(img_height, img_width))
rgb_img = skimage.color.gray2rgb(img)
imagenet_height, imagenet_width = 224, 224
seq = iaa.Sequential(
[iaa.Scale({
"height": imagenet_height,
"weight": imagenet_width
})])
seq_det = seq.to_deterministic()
image_aug = seq_det.augment_images([rgb_img])[0]
bbs_aug = seq_det.augment_bounding_boxes([bb_array])[0]
return image_aug, bbs_aug
def to_imgaug(self, shape):
"""
Args:
shape (tuple): shape of image that boxes belong to
Example:
>>> self = Boxes([[25, 30, 15, 10]], 'tlbr')
>>> bboi = self.to_imgaug((10, 10))
"""
import imgaug
if len(self.data.shape) != 2:
raise ValueError('data must be 2d got {}d'.format(len(self.data.shape)))
tlbr = self.to_tlbr(copy=False).data
bboi = imgaug.BoundingBoxesOnImage(
[imgaug.BoundingBox(x1, y1, x2, y2)
for x1, y1, x2, y2 in tlbr], shape=shape)
return bboi
def test_to_batch_in_augmentation__all_columns(self):
batch = ia.Batch(
images=np.zeros((1, 2, 2, 3), dtype=np.uint8),
heatmaps=[
ia.HeatmapsOnImage(np.zeros((2, 2, 1), dtype=np.float32),
shape=(2, 2, 3))
],
segmentation_maps=[
ia.SegmentationMapsOnImage(np.zeros((2, 2, 1), dtype=np.int32),
shape=(2, 2, 3))
],
keypoints=[
ia.KeypointsOnImage([ia.Keypoint(x=0, y=0)], shape=(2, 2, 3))
],
bounding_boxes=[
ia.BoundingBoxesOnImage([ia.BoundingBox(0, 0, 1, 1)],
shape=(2, 2, 3))
],
polygons=[
ia.PolygonsOnImage([ia.Polygon([(0, 0), (1, 0), (1, 1)])],
shape=(2, 2, 3))
],
line_strings=[
ia.LineStringsOnImage([ia.LineString([(0, 0), (1, 0)])],
shape=(2, 2, 3))
])
batch_inaug = batch.to_batch_in_augmentation()
assert isinstance(batch_inaug, ia.BatchInAugmentation)
assert ia.is_np_array(batch_inaug.images)
assert batch_inaug.images.shape == (1, 2, 2, 3)
assert isinstance(batch_inaug.heatmaps[0], ia.HeatmapsOnImage)
assert isinstance(batch_inaug.segmentation_maps[0],
ia.SegmentationMapsOnImage)
assert isinstance(batch_inaug.keypoints[0], ia.KeypointsOnImage)
assert isinstance(batch_inaug.bounding_boxes[0],
ia.BoundingBoxesOnImage)
assert isinstance(batch_inaug.polygons[0], ia.PolygonsOnImage)
assert isinstance(batch_inaug.line_strings[0], ia.LineStringsOnImage)
assert batch_inaug.get_column_names() == [
"images", "heatmaps", "segmentation_maps", "keypoints",
"bounding_boxes", "polygons", "line_strings"
]
def apply_augmenters(self, image, target):
"""
Args:
image: image to apply augmentation
target: target to apply augmentation
"""
if self.seq is None:
target = np.array(target, dtype=np.float32)
target[:, 0] /= image.shape[1]
target[:, 1] /= image.shape[0]
target[:, 2] /= image.shape[1]
target[:, 3] /= image.shape[0]
else:
seq_det = self.seq.to_deterministic()
image = seq_det.augment_images([image])[0]
boxes = ia.BoundingBoxesOnImage([
ia.BoundingBox(x1=target[i][0],
y1=target[i][1],
x2=target[i][2],
y2=target[i][3])
for i in range(target.shape[0])
],
shape=image.shape)
boxes = seq_det.augment_bounding_boxes([boxes])[0]
image_after = boxes.draw_on_image(image,
thickness=2,
color=[0, 0, 255])
target = np.array([[
box_aug.x1 / image.shape[1], box_aug.y1 /
image.shape[0], box_aug.x2 / image.shape[1],
box_aug.y2 / image.shape[0], target[i][-1]
] for i, box_aug in enumerate(boxes.bounding_boxes)],
dtype=np.float32)
if self.is_BGR:
image = image[..., ::-1]
target[:, :4] = target[:, :4].clip(min=0.0, max=1.0)
image = cv2.resize(image, (self.image_size, self.image_size))
return image, target
def make_augumentation(image_coords, seq=None):
image = cv.imread('{}/{}'.format(LOAD_IMGS_PATH, image_coords[0]))
points = []
for i in image_coords[1]:
points.append(ia.BoundingBox(i[0], i[1], i[2], i[3]))
bbs = ia.BoundingBoxesOnImage(points, shape=image.shape)
seq_det = seq.to_deterministic()
image_aug = seq_det.augment_images([image])[0]
bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
# print coordinates before/after augmentation (see below)
# use after.x_int and after.y_int to get rounded integer coordinates
points_aug = []
h, w = image.shape[:2]
# print( '{} {}'.format( w, h ) )
# print( image_coords[0] )
for i in range(len(bbs.bounding_boxes)):
after = bbs_aug.bounding_boxes[i]
x1 = after.x1_int
y1 = after.y1_int
x2 = after.x2_int
y2 = after.y2_int
x1 = 0 if x1 < 0 else (w - 1 if x1 >= w else x1)
y1 = 0 if y1 < 0 else (h - 1 if y1 >= h else y1)
x2 = 0 if x2 < 0 else (w - 1 if x2 >= w else x2)
y1 = 0 if y2 < 0 else (h - 1 if y2 >= h else y2)
if (x1 < 0 or x1 >= w or x2 < 0 or x2 >= w or y1 < 0 or y1 >= h
or y2 < 0 or y2 >= h):
raise ValueError('Has incorrect boundings.')
points_aug.append([x1, y1, x2, y2])
image_after = bbs_aug.draw_on_image(image_aug,
thickness=2,
color=[0, 0, 255])
return image_after, image_aug, points_aug
def parser_label(self, image, yolo_label):
label = []
for h_index in range(self.cell_size):
for w_index in range(self.cell_size):
if yolo_label[h_index, w_index, 0] == 0:
continue
x_center, y_center, w, h = yolo_label[h_index, w_index, 1:5]
class_id = np.argmax(yolo_label[h_index, w_index, 5:])
x_1 = int((x_center - 0.5 * w) * self.image_size)
y_1 = int((y_center - 0.5 * h) * self.image_size)
x_2 = int((x_center + 0.5 * w) * self.image_size)
y_2 = int((y_center + 0.5 * h) * self.image_size)
label.append(
ia.BoundingBox(x1=x_1,
y1=y_1,
x2=x_2,
y2=y_2,
label=class_id))
return image, ia.BoundingBoxesOnImage(label, shape=image.shape)
def get_crop_dataset(augmenter, image, bboxs, max_h, max_w):
temp_aug_bbox = []
for bbox in bboxs:
temp_aug_bbox.append(ia.BoundingBox(x1=bbox[0],
x2=bbox[2],
y1=bbox[1],
y2=bbox[3]))
bbs = ia.BoundingBoxesOnImage(temp_aug_bbox, shape=image.shape)
# keep same aug
seq_det = augmenter.to_deterministic()
res_img = seq_det.augment_image(image)
bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
res_bboxs = []
for bbx in bbs_aug:
if bbx.x1 >= 0 and bbx.y1 >= 0 and bbx.x2 < max_w and bbx.y2 < max_h:
res_bboxs.append([int(bbx.x1), int(bbx.y1), int(bbx.x2), int(bbx.y2)])
return res_img, res_bboxs
def augment_flip(image, bbox):
aug = iaa.Sequential([iaa.Fliplr(1.0)])
bbs = ia.BoundingBoxesOnImage([
ia.BoundingBox(x1=bbox[0], y1=bbox[1], x2=bbox[2], y2=bbox[3])], shape=image.shape)
aug = aug.to_deterministic()
image_aug = aug.augment_image(image)
image_aug = image_aug.copy()
bbs_aug = aug.augment_bounding_boxes([bbs])[0]
b = bbs_aug.bounding_boxes
bbs_aug = [b[0].x1, b[0].y1, b[0].x2, b[0].y2]
bbs_aug = np.asarray(bbs_aug)
bbs_aug[0] = bbs_aug[0] if bbs_aug[0] > 0 else 0
bbs_aug[1] = bbs_aug[1] if bbs_aug[1] > 0 else 0
bbs_aug[2] = bbs_aug[2] if bbs_aug[2] < size else size
bbs_aug[3] = bbs_aug[3] if bbs_aug[3] < size else size
return image_aug, bbs_aug
def get_bounding_boxes(self, width: int,
height: int) -> list[ia.BoundingBox]:
bounding_boxes = []
for group in self.keypoint_groups:
group_x = [k.x for k in group]
min_x = max(0, int(min(group_x) - BOUNDING_BOX_BUFFER))
max_x = min(width, int(max(group_x) + BOUNDING_BOX_BUFFER))
group_y = [k.y for k in group]
min_y = max(0, int(min(group_y) - BOUNDING_BOX_BUFFER))
max_y = min(height, int(max(group_y) + BOUNDING_BOX_BUFFER))
bounding_boxes.append(
ia.BoundingBox(x1=min_x,
y1=min_y,
x2=max_x,
y2=max_y,
label=self.name))
return bounding_boxes
def loadData(imgPath, txtPath):
outBoxes = []
tmpBoxes = []
img = cv2.imread(imgPath)
h, w , c = img.shape
outImg = img.reshape(1, h, w, c)
with open(txtPath, "r") as f:
for line in f.readlines():
bbox = line.strip().split(" ")
label = bbox[0]
cx, cy, bw, bh = float(bbox[1]) * w, float(bbox[2]) * h, float(bbox[3]) * w, float(bbox[4]) * h
x1, y1 = cx - 0.5 * bw, cy - 0.5 * bh
x2, y2 = cx + 0.5 * bw, cy + 0.5 * bh
tmpBoxes.append(ia.BoundingBox(x1, y1, x2, y2, label))
outBoxes.append(tmpBoxes)
return outImg, outBoxes
def augmentation(images, annotations, augmentation, save_dir, filename):
# print("\n", augmentation, "\n")
for idx in range(len(images)):
image = images[idx]
boxes = annotations[idx][0]
ia_bounding_boxes = []
for box in boxes:
xyxy = xywh2xyxy(box, Width, Height)
# print(xyxy)
ia_bounding_boxes.append(
ia.BoundingBox(x1=xyxy[0],
y1=xyxy[1],
x2=xyxy[2],
y2=xyxy[3],
label=box[0]))
bbs = ia.BoundingBoxesOnImage(ia_bounding_boxes, shape=image.shape)
seq = iaa.Sequential(augmentation)
seq_det = seq.to_deterministic()
image_aug = seq_det.augment_images([image])[0]
bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
new_image_file = save_dir + filename + annotations[idx][2]
cv2.imwrite(new_image_file, image_aug)
h, w = np.shape(image_aug)[0:2]
# print("h, w : ", h, w)
# voc_writer = Writer(new_image_file, w, h)
with open(save_dir + filename + annotations[idx][1], "w") as f:
for i in range(len(bbs_aug.bounding_boxes)):
bb_box = bbs_aug.bounding_boxes[i]
x_min = bb_box.x1
y_min = bb_box.y1
x_max = bb_box.x2
y_max = bb_box.y2
cls_id = bb_box.label
x_cen, y_cen, w, h = xyxy2xywh(x_min, y_min, x_max, y_max)
f.write("%d %.06f %.06f %.06f %.06f\n" %
(cls_id, x_cen, y_cen, w, h))
def test_two_images_and_bounding_boxes(self):
rng = iarandom.RNG(0)
images = rng.integers(0, 256, size=(2, 256, 256, 3), dtype=np.uint8)
bbs = []
for x in np.linspace(0, 256, 5):
for y in np.linspace(0, 256, 5):
bbs.append(ia.BoundingBox(x1=x, y1=y, x2=x + 20, y2=y + 20))
bbsoi1 = ia.BoundingBoxesOnImage(bbs, shape=images[0].shape)
bbsoi2 = bbsoi1.shift(left=20)
image1_w_overlay = bbsoi1.draw_on_image(images[0])
image2_w_overlay = bbsoi2.draw_on_image(images[1])
debug_image = iaa.draw_debug_image(images,
bounding_boxes=[bbsoi1, bbsoi2])
assert self._image_contains(images[0, ...], debug_image)
assert self._image_contains(images[1, ...], debug_image)
assert self._image_contains(image1_w_overlay, debug_image)
assert self._image_contains(image2_w_overlay, debug_image)
def pull_item(self, index):
img_id = os.path.join(self.image_path, self.images[index])
boxes = self.boxes[index]
labels = np.ones(len(boxes)) * 0
if len(boxes) == 0:
return None, None, None
img = io.imread(img_id)
box = np.asarray(boxes)
if True:
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
img, self.image_size)
if len(labels) == 0:
return ig, box, labels
box = box * scale
box[:, 0] = box[:, 0] + padding[1][0]
box[:, 1] = box[:, 1] + padding[0][0]
box[:, 2] = box[:, 2] + padding[1][1]
box[:, 3] = box[:, 3] + padding[0][1]
bb = []
for ix, x in enumerate(box):
bb.append(ia.BoundingBox(x[0], x[1], x[2], x[3], labels[ix]))
bbs = ia.BoundingBoxesOnImage(bb, shape=self.image_size)
seq_det = self.img_aug.to_deterministic()
image_aug = seq_det.augment_images([ig])[0]
bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
box = []
labels = []
for i in range(len(bbs.bounding_boxes)):
after = bbs_aug.bounding_boxes[i]
box.append([after.x1, after.y1, after.x2, after.y2])
labels.append(after.label)
box = np.asarray(box)
box = box / np.asarray([
self.image_size[1], self.image_size[0], self.image_size[1],
self.image_size[0]
])
box = np.clip(box, 0, 1)
return image_aug, box, labels
def aug_image(self, train_instance, categories, jitter):
image_ann, annotations = train_instance
image = cv2.imread(image_ann['file_name'])
h, w = image.shape[:2]
categories = {
category['id']: category['name']
for category in categories
}
if image is None:
print('Cannot find ', image['file_name'])
aug_pipe_deterministic = self.aug_pipe.to_deterministic()
all_objs = [{
'xmin': int(w * x['bbox'][0][0]),
'ymin': int(h * x['bbox'][0][1]),
'xmax': int(w * x['bbox'][1][0]),
'ymax': int(h * x['bbox'][1][1]),
'name': categories[x['category_id']]
} for x in annotations if x['category_id'] in categories]
bbs = ia.BoundingBoxesOnImage([
ia.BoundingBox(x1=obj['xmin'],
y1=obj['ymin'],
x2=obj['xmax'],
y2=obj['ymax'],
name=obj['name']) for obj in all_objs
],
shape=image.shape)
image = cv2.resize(image,
(self.config['IMAGE_H'], self.config['IMAGE_W']))
image = np.copy(image)
bbs = bbs.on(image)
if jitter:
image = aug_pipe_deterministic.augment_image(image)
bbs = aug_pipe_deterministic.augment_bounding_boxes([bbs])[0] \
.cut_out_of_image().remove_out_of_image()
return image, bbs.bounding_boxes
